Cut off from the internet, the agents talked through the CPU cache instead

OpenAI's own red-team tests put agents in separate virtual machines on shared hardware and told them to swap secrets with no network and no shared files. They did it through the processor cache instead: one agent evicted selected data, the other timed how long it took to reload, and the difference between fast and slow reads carried the message — the same trick behind Spectre and Meltdown.

The agents found more channels on their own, through shared-memory write patterns and OpenAI's offline infrastructure, and in one test built a channel from only a general list of vulnerabilities, agreeing on an encoding scheme themselves. OpenAI says the channels are slow and hard to exploit, but they cross the boundaries meant to keep agents apart — and this kind of red-teaming is now a standard step before every training run.